Tk Breast (1994) 3,222-226 Q Longman
GroupLtd
1994
THE
BREAST
I Work in Progress
I
Classification of equivocal mammograms through digital analysis K. S. Coakley, F. Quintarelli,
T. van Doorn and C. Hirst*
Centre for Medical and Health Physics, Queensland Hospital, Brisbane, Australia
University of Technology,
“Breast Screening
Clinic, Wesley
S U M M A R Y. In a clinical trial, 45 equivocal mammograms were evaluated using digital image processing techniques and the results were correlated with pathology reports. 31 cases agreed with the histological diagnosis, 10 of the other 14 cases were false negatives and 4 false positive (sensitivity 33%, specificity 87%). Factors contributing to difficulties in analysis included: (a) low density calcifications, (b) high density background parenchyma and (c) tightly packed clusters of microcalcification. This work has shown that digital analysis can reduce the number of equivocal mammographic examinations and potentially decreases the need for examination by biopsy, however, many technical problems remain before this technology can be applied to all such clusters.
INTRODUCTION
are such that the cluster can be classified only as indeterminate or equivocal. In these cases, the patient and her doctor are faced with an unsatisfactory choice between the need for histological or cytological sampling by means of needle or open biopsy or alternatively, shortterm mammographic review (usually 6 months). If computer analysis could resolve the dilemma of diagnosis of these clusters of calcification, the diagnostic process would be enhanced without the need for invasive procedures and without the risk of delaying the diagnosis. This paper describes a method of improving the predictability of malignancy of mammographic calcification. Several authors have studied breast cancer through mammography qualitatively46 and some work has been carried out on identification of calcification patterns by digital analysis.7 As early as the 1970s analysis of clusters by computer enjoyed some success.8 However, the difference between these studies and that described in this paper is that they evaluate all classes of clusters of calcification, whereas this study looks only at equivocal or indeterminate clusters. Other successful studies have considered computer classification of mammograms with neural networks. g~loPatrick et al” used expert learning system networks on image and clinical data to recognize 60% of benign clusters and 90% of malignant cluster. Whilst radiological assessment of microcalcification is usually carried out using both crania-caudal and vertical images, this study has examined the clusters in crania-caudal images only. The authors considered that if discrimination were possible using one view only then the complexity of the computer analysis would be significantly reduced. Analysis of clusters is based on
Breast cancer is the most common form of cancer occurring in women, with nearly 70% of breast cancers occurring in women over the age of 50.’ Studies have demonstrated that mammographic screening can reduce the number of deaths from breast cancer by as much a s 30% in this age group.’ Breast cancer is an abnormal proliferation of epithelial cells in the ducts or lobules of the breast.2 Ductal-carcinoma-in-situ (DCIS) is an early stage of breast cancer in which the malignant process is confined within ducts and lobules. By definition, metastatic spread should not be possible and removal of DCIS by surgical excision should, theoretically, cure the patient.3 Mammographic microcalcification is the predominant means of identification of DCIS. Microcalcifications are deposits of calcium salts in tissues. They resemble grains of salt and take various shapes - mostly punctate or elongated. Clusters of calcifications may occur within ductal structure’s within or in association with a malignant proliferative process or alternatively they may be generated by a number of benign processes. The deposits of calcium vary in all parameters but particularly size, shape and density. In many instances, a film reader can diagnose the nature of a pathophysiological process by virtue of the morphology of the calcifications. However, in a significant proportion (3040%) the features of the calcific flecks
Address correspondence to: K. S. Coakley, Centre for Medical and Health Physics, Queensland University of Technology, GPO Box 2434, Brisbane, Queensland, 4001, Australia
222
E$~ivocal mammograms
through digital analysis
223
t.8 magnification view using a localized compression technique. In all instances, GEC Senographe @JOT mammography units have been used to generate the images on Kodak film with dedicated processing using Kodak chemktry. Each cluster of calcification is assessed with regard to the probability of malignancy using a widely accepted classification system. The grading is based on morphological features and distribution within the breast. The following classification of grade was used:
a x
1. 2. 3. 4. 5.
no abnormality is detected benign equivocal/probably benign suspicious malignant.
UsuaI practice is that lesions graded (2) will be treated conservatively, those graded (4) or (5) excised, whereas lesions graded (3) may be subjected to cytology, core biopsy, open biopsy or alternatively, reviewed. This work examines clusters of micro-calcification which were graded as equivocal. Clusters of calcification which were clearly diagnosable either as benign or malignant by the film reader were not included within the study. To evaluate the accuracy of the technique, the computer-based analyses were correlated with biopsy results. ‘Through a process of enhancement and analysis of calcification patterns it was hoped that it would be possible to discriminate between benign and malignant processes.
Within the defined parameters of the study, the image to be assessed by computer analysis consisted of two main features, a cluster of calcification and background tissue. Isolated large calcifications usually indicate a benign process and tend to be ignored in assessment of the calcification by experienced film readers. Smaller flecks isolated from the main cluster are also usually discounted in the analysis. Attention is focussed on the smaller calcifications and their specific morphological characteristics. Background tissue varies greatly in density from one woman to another and will be dependent on the position of the calcification within the breast. Within the analyzed mammographic image there may be marked variation in density of the background parenchyma, or alternatively, marked or minimal contrast between the calcification and background tissue. The difference in optical density between the calcification and tbe background plays a large part in the image processing techniques and contributes to some of the difficulties encountered in the analysis.
Square regions (5 cm by 5 cm) of magnified (x 1.8) crania-caudal views were digitized on a Lumisys
Fig. l-Digitized region of an equivocal mammogram demonstrating the range of size, shape and distribution patterns OFcakifications.
kumiscan 100 flat bed digitizer to a 12 bit density resolution and IO0 pm spatial resohttion. Each pixel equated to 54 ~-lm sampling at the exit side of the breast. The digital image was then manipulated using a SILICON GRAPHICS Power Series 4D/21O/GTX workstation and displayed using the ANALYZE package.12 Figure I shows a typical digitised image.
Image
preprocessing
Unwanted background soft tissue, noise, large calcifications and isolated calcifications were removed with an image preprocessing algorithm (Fig. 2). Stage 1: Elimination of soft tissue. Soft tissue structures are relatively large and can be eliminated through high pass filtering (gaussian cross section notch filter with a dc value of 0.7 and a bandwidth of 200:). The calcifications of interest are predominantly small and unaffected by the filtering process. This image was then low pass filtered (standard gaussian with a dc value of unity and a bandwidth of 100) to eliminate the quantum mottle.13 The filter removed noise that corresponds to a scale size of approximately one pixel. A binary image was formed with an experimentally determined global threshold. The binary image consisted of small high intensity regions corresponding to medium or large calcifications. Extralarge calcifications fragmented by the high pass filtering appeared as tight clusters. Stage 2: Isolation of :extra-large calci$catiorzs. The following process generated an image collsisting only of extra-large calcifications. A high pass filter with a dc component of 0.3 and bandwidth of 20 was applied to the original image. This lightly filtered image was then thresholded and morphologically fi1teredi”‘6 with an open function followed by a close function leaving an image containing only extra-large calcifications at their original size and shape.
224
The Breast
cification clusters comprise a main cluster with a few calcifications separated from the main group. In radiographic terms these peripheral calcifications are of no interest. Isolated calcifications were removed by calculating the distance between each calcification and its three closest neighbours; a calcification was excluded from the cluster if there were < 2.5 calcifications/cm2.
Identification 1 THRESHOLD ( I
INVERSE FOURIER TRANFORM THRESHOLD TO FORM BINARY IMAGE I
I
IMAGE OF LARGE CALCIFICATIONS
of calcifications
A further series of algorithms was developed to compile a data set of identifying parameters for each calcification. An image was scanned line by line. The detection of a calcification boundary resulted in the calculation of the following parameters: coordinates of centre area optical density perimeter. The algorithm subsequently removed the calcification and then searched for the next.
” Classification
PREPROCESSEDIMAGE Fig. &image
preprocessing
of calcification
patterns
The calcifications were sorted into classifications according to their size, density and shape parameters.
flowchart.
Stage 3: Removal of extra-large calcifications. The outcome of previous image preprocessing stages was two binary images. One retained all calcifications in the original image with fragmented extra-large calcifications and the other contained only the extra-large calciftcations. The difference between the two images was an image that contained solely the calcifications of interest (Fig. 3).
Size. Calcifications were sorted into four area categories - small, medium, large and extra large. The small category was for calcifications of less than 6 pixels (0.017 mm2) in area. The medium category included calcifications of between 7 and 29 pixels (0.02 m2 and 0.085 mm’). Large describes all calcifications of size 30-300 pixels (0.088-0.88 mm*). The fourth category, extra large, was for the extra large calcifications previously removed by the morphological filtering.
Stage 4: Removal of isolated calcifications. Most calDensity. Each calcification was sorted into one of three groups according to its optical density in the original digitized image. Cut-off values for the three groupings were empirically determined and set as follows: low density - greater than the mean minus three standarc 1J deviations high density - less than the mean minus four standarr Ll deviations medium density - all other calcifications. The use of relative values for the classification of density obviated the need for constant X-ray exposure factors; assuming the mammograms were exposed within the linear region of the response curve of the film/screen. Shape. A shape descriptori
of the form:
SHAPE DESCRIPTOR = (Perimeter)*/4nArea Fig. L-Preprocessed binary image showing the calcifications interest along with the fragmented extra large calcifications.
of
was used to categorize the calcifications
as either circular
Equivocal
or elongated. A shape descriptor ranging from 0.75 to 1.2 was set for a circular classification. The shape descriptor has a minimum value of unity for a circle and greater than unity otherwise. However, when dealing with very small calcifications, the spatial quantization of the image (i.e. pixel size) gives rise to values of the shape descriptor that are less than unity. Any shape determination ignored very small calcifications because of their unreliability.
CIassIfica~ion of calcification characteristics basis of histologic correlation
as a
A model of assessment of the morphology of the calcific flecks used by experienced film readers was used as a basis for the computer analysis of the calcification. An initial categorization by size into three groups was made, as follows: Group 1 - clusters containing > 25% small and > 20% large flecks Group 2 - clusters witt > 25% small and < 20% large flecks Group 3 - < 2590 small flecks. The second stage of analysis was to consider within each group, different percentages of variation in shape and density.
RESIJLTS The results of a computer analysis of 45 clusters of microcalcification identified on mammography were correlated with the relevant biopsy reports (Table). The first column in the Table presents the grading system that we are using to analyze the clusters. Column 2 shows the number of clusters classified into each grading by the digital analysis system. Columns 3 and 4 contain. the number of clusters classified by biopsy results as benign or malignant. With a cutoff for a malignant classification of the cluster set to grades med-high, high and undefined, a sensitivity (percentage malignant biopsy with a ‘positive’ test) of 33% and a specificity (percentage benign biopsy with test ‘negative’) of 87% was achieved. If the cutoff for malignancy was set to all grades above med-low, the sensitivity increased to 60% but the specificity was reduced (67%).
mammograms
through digital analysis
225
The outcome of the study, is encouraging although the sensitivity and specificity do require improvement. The image processing algorithms successfully identified the calcifications of interest. The analysis algorithms appropriately classified a high proportion of calcifications in agreement with the biopsy results. Of the 45 mammograms, 31 were appropriately identified, IO were false negatives and 4 false positive results. Clearly there is room for improvement in the accurate prediction of histological process involved. Three main problem areas were identified in the process of analysis: 8 Low density calcific flecks 8 Densely clustered calcifications @ A very dark mammographic background. Very low density calcific flecks (as shown in Figure 4) present the greatest difficulty in the successful classification of malignant calcification patterns. When the computer program pre-processes these low density sections, one of two events can occurs: the low density shape is broken down into an entirely different shape or it is disregarded. In future work it is hoped that this will be overcome by analyzing the density first., Densely clustered flecks occur in both malignant and benign cases. The difficulty with these clusters is that the algorithm can go into an infinite figure of eight loop around two pixels when attempting to determine a calcification contour. Very dark images give incorrect or null results because the dark tissue is interpreted by the algorithm as a huge calcification. The null result is due to the contour routine because the dark tissue is often at the edge of the image and the algorithm loops continuously. An incorrect result occurs because a huge calcification is always interpreted as benign. Elimination of these dark images will require an improvement in the technical factors in the production of the initial mammogram. A number of other technical pro&ems occur when the
T&Be Table of digital analysis and biopsy results for all patients in the clinical trial Grading
Low Medium-low Medium Medium-high High Undefined
Biopsy result
Number classified by digital analysis
Benign
2 24 10 1 0 8
I
1
19 6 0 0 4
5 4
Malignant
I 0 4
Fig. &Computer magnified region of an image showing an example of a low grade DCIS. The calcifications are of similar density to the background tissue.
226
The Breast
image includes vessels or ducts full of necrotic material. Using the present algorithm, it is not possible to analyze any clusters that incorporate vessels. The calcifications need to be enhanced with a different filtering process while eliminating the vessel. The determination of the onset of breast malignancy by means of mammographic calcification is not absolute. A significant proportion of clustered microcalcifications cannot be identified as either benign or malignant by visual inspection of the mammogram. This research program has attempted to further the differentiation of equivocal mammograms through computer enhancement and analysis. Although a high proportion of the mammograms were successfully classified, further work is required to improve the sensitivity and specificity. The most important factors to be overcome with the computer analysis is the segmentation of low density calcifications, calcifications in dense clusters and calcifications in dense soft tissue. Acknowledgements The authors wish to express their gratitude to the Centre for Medical and Health Physics, Queensland University of Technology for the use of digitizing and computing resources, the Wesley Hospital for the clinical data sets, and J. Pfitzner for assistance in the implementation of the algorithms.
References 1. Information kit on breast cancer screening programs; Women’s Cancer Prevention Program. Queensland Health Department 1993. 2. Page D, Anderson T. Diagnostic histopathology of the breast. Edinburgh: Churchill Livingstone, 1987.
3. Lanyi M. Diagnosis and differential diagnosis of breast calcifications. Springer-Verlag, 1988. 4. Wolfe J, Albert S, Belle S, &lane M. Breast parenchymal patterns and their relationship to risk for having or developing carcinoma. Radio1 Clin North Am 1983; 21: 127-136. 5. Weaver M. Breast cancer in nonpalpable lesions: can mammographic parenchymal pattern improve the predictive value of biopsy? Am Surg 1992; 58: 692-694. 6. Albertyn L E. Mammographically indeterminate microcalcifications - can we do any better? Au&alas Radio1 1991; 35: 350-357. 7. Magnin I E, El Alaoui M, Bremond A. Automatic microcalcifications pattern recognition from X-ray mammography. SPJE 1989; 1137: 170-175. 8. Wee W, Moskowitz M, Chang N, Ting Y, Pemmeraju S. Evaluation of mammographic calcifications using a computer program. Radiology 1975; 116: 717-720. 9. Wu Y, Geiger M, Doi K, Vybomy C, Schmidt R, Metz C. Artificial neural networks in mammography: application to decision making in the diagnosis of breast cancer. Radiology 1993; 187: X1-87. 10. Getty D, Pickett R, D’Orsi C, Swets J. Enhanced interpretation of diagnostic images. lnvestigative Radiology 1988; 23: 240-252. 11. Patrick E A, Moskowitz M, Mansukhani V T, Gruenstein E I. Expert learning system network for diagnosis of breast calcifications. Invest Radio1 1991; 26: 534-539. 12. Robb R A. A software system for interactive and quantitative visualisation of multidimensional biomedical images. Aust Phys Eng Sci Med, 1991; 14: 9-30. 13. Castleman K. Digital image processing. Prentice-Hall, 1979. 14. Jain A. Fundamentals of digital image processing. Prentice-Hall, 1989. 15. Gonzalez R, Wintz P. Digital image processing. Addison Wesley, 1977. 16. Vogt R. Automatic generation of morphological set recognition algorithms. Spring-Verlag, 1989. 17. Davies D H, Dance D R, Jones C H. Automatic detection of microcalcifications in digital mammograms using local area thresholding techniques. SPIE 1989; 2: 1092.
Date received 9 August 1993 Date accepted 24 March 1994